Network files (specified by the -F command line parameter) contain the
representation learned by SNoW for each of the target nodes as well as all the
information required for SNoW to recreate (in testing) the structure learned
during training. Each target node has a section in the file containing
information on its algorithm's parameters and features (most notably, the
features' weights). The header line for each target node takes the following
form:
target ID priorProbability cloudConfidence activeCount nonActiveCount
algorithm
learnerType parameters
For example:
target 2 0.4 0.473593433165 42 63 winnow 1 1.35 0.8 4 0.2
This specifies a target node with ID 
, prior probability 
(the
fraction of training examples labeled with the target ID), and cloud
confidence 
(see Section 4.3.5), which appeared active in
examples, and was not found active in 
examples. It uses the Winnow
algorithm found at index 
of an internal data structure that maintains
information on every algorithm used by the network and their parameters. The
Winnow algorithm used by this target node has an 
of 
, a 
of 
, a threshold of 
, and a default feature weight of 
.
Following the target node header, the target node's features are enumerated.
Each line corresponds to a single feature, in the format:
ID : learnerType : featureID : activeCount updates weight
For example:
1 : 2 : 34 : 13 6 0.3645
This specifies a feature which is linked to target node using algorithm
, and its ID is 
. It had an active count of 
, and its weight was
updated 
times during training. After those updates, its weight is now
.
For those users who are concerned about a feature with very large ID found in their network files, take a look at the -f command line parameter.